I ran an AI nuclear engineering department for a week (charlesazam.com)

In an innovative experiment, an AI researcher tasked autonomous agents with solving three complex nuclear engineering challenges over a week. Using a €30/month server and a relatively modest $125 in API costs, the agents tackled problems that involved modeling a passive cooling rig, predicting fuel failure rates, and simulating a reactor's self-rescue capabilities. The agents demonstrated remarkable accuracy, closely matching real-world measurements in completed tasks while also revealing the pitfalls in existing nuclear industry models. This experiment is significant for the AI and machine learning community as it underscores the potential of advanced AI systems to engage deeply with complex physical systems, effectively learning and generating insights from provided data. The results revealed that the agents could reproduce systematic errors inherent in the nuclear field, identified gaps in the provided physics information, and even coded new physics terms when needed. Challenges faced, such as varying prediction quality based on the completeness of input data, illustrate the importance of comprehensive datasets for reliable model performance. Overall, the study highlights both the capabilities and limitations of AI in engineering contexts, paving the way for future advancements in automated simulations and predictive analytics.